Where interactions live: silencing interactions in two classes of fly brain model need a strong member, and the battery that follows certifies most but not all combinations

Papers 13 and 14 found that silencing interactions in connectome-constrained models of the fly concentrate on a few cell types with strong single effects, but neither measured the form of that rule nor had a complete spectrum to test it on. Here the rule was tested as heredity, a term from statistical interaction models: in the strong form an interaction above tolerance needs every member to have a strong single effect; in the weak form it needs at least one. Complete interaction spectra (every subset of an 11-type universe, 2,047 sets, exact Möbius inversion through order 11) were measured in all 50 task-trained flyvis models of the visual system (two readouts, 204,700 items) and in the untrained whole-brain spiking model of Shiu et al. (2024) at two sugar drives (2 × 2,168 sets, 189,280 trials), the second chosen so that two types saturate the readout. Each universe mixed the eight strongest types with three weak ones. Five hypotheses were frozen before the first confirmatory simulation and the analysis was run once. Weak heredity held: across flyvis models, at most 0.78% of above-tolerance interactions at any order sat in sets with no strong member, and none at orders 6-10; at the saturating Shiu drive, none did. The strong form failed, as preregistered: the share of above-tolerance interactions confined to all-strong sets fell from 0.84 at pairs to 0.23 at order 6. Sets of weak types alone were additive at orders 4-6 but not quite at pairs and triples (1.7% and 1.1% above tolerance against a 1% threshold). Saturation did not explain where interactions live: clip-touching was enriched among above-tolerance mixed sets only at orders 3-8, and at the saturating Shiu drive interaction energy rose among strong types (pair energy 0.48 against 0.02 at the non-saturating drive) but stayed indistinguishable from zero in mixed sets. The battery implied by heredity (every single, every combination of strong types, weak members added) beat the additive model by 0.22-0.35 at every order in flyvis and 0.27-0.49 at the saturating drive, but reached the preregistered 85% within-tolerance bar only at order 3 in flyvis (0.874, then 0.847, 0.832 and 0.822 at orders 4-6). A sparse estimator built on the heredity prior did not beat simple screens at this universe size, largely as a synthetic preview had predicted. An independent re-derivation from the raw outputs reproduced every count and point estimate.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24330061
Primary Topic
Neurobiology and Insect Physiology Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Where interactions live: silencing interactions in two classes of fly brain model need a strong member, and the battery that follows certifies most but not all combinations

Thomas Ryan
Zenodo (CERN European Organization for Nuclear Research)
Neurobiology and Insect Physiology Research
preprint

Where interactions live: silencing interactions in two classes of fly brain model need a strong member, and the battery that follows certifies most but not all combinations

Thomas Ryan
preprint en

Abstract

Papers 13 and 14 found that silencing interactions in connectome-constrained models of the fly concentrate on a few cell types with strong single effects, but neither measured the form of that rule nor had a complete spectrum to test it on. Here the rule was tested as heredity, a term from statistical interaction models: in the strong form an interaction above tolerance needs every member to have a strong single effect; in the weak form it needs at least one. Complete interaction spectra (every subset of an 11-type universe, 2,047 sets, exact Möbius inversion through order 11) were measured in all 50 task-trained flyvis models of the visual system (two readouts, 204,700 items) and in the untrained whole-brain spiking model of Shiu et al. (2024) at two sugar drives (2 × 2,168 sets, 189,280 trials), the second chosen so that two types saturate the readout. Each universe mixed the eight strongest types with three weak ones. Five hypotheses were frozen before the first confirmatory simulation and the analysis was run once. Weak heredity held: across flyvis models, at most 0.78% of above-tolerance interactions at any order sat in sets with no strong member, and none at orders 6-10; at the saturating Shiu drive, none did. The strong form failed, as preregistered: the share of above-tolerance interactions confined to all-strong sets fell from 0.84 at pairs to 0.23 at order 6. Sets of weak types alone were additive at orders 4-6 but not quite at pairs and triples (1.7% and 1.1% above tolerance against a 1% threshold). Saturation did not explain where interactions live: clip-touching was enriched among above-tolerance mixed sets only at orders 3-8, and at the saturating Shiu drive interaction energy rose among strong types (pair energy 0.48 against 0.02 at the non-saturating drive) but stayed indistinguishable from zero in mixed sets. The battery implied by heredity (every single, every combination of strong types, weak members added) beat the additive model by 0.22-0.35 at every order in flyvis and 0.27-0.49 at the saturating drive, but reached the preregistered 85% within-tolerance bar only at order 3 in flyvis (0.874, then 0.847, 0.832 and 0.822 at orders 4-6). A sparse estimator built on the heredity prior did not beat simple screens at this universe size, largely as a synthetic preview had predicted. An independent re-derivation from the raw outputs reproduced every count and point estimate.

Zenodo (CERN European Organization for Nuclear Research)
Neurobiology and Insect Physiology Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon