Two models of one zebrafish connectome had different stability under nominal 10% gain CV with fixed mean gain and initial 10 s leading decay

We tested how neuron gain changes affect stability in a rate model from one larval zebrafish brainstem connectome. We kept 289 core cells and changed gains in 255 integrator cells. Two common-time models had the same initial 10 s leading decay. One used a 1 s intrinsic time and recurrent eigenvalue 0.9. The other used 0.1 s and 0.99. We drew positive lognormal gains with 10% coefficient of variation, then set their sample mean to one. We kept recurrent scale fixed. The 0.1 s model was unstable in 212 of 1,000 draws. The 1 s model was stable in all 1,000 draws. Monte Carlo error ranges were 18.8–23.8% and 0–0.38%, respectively. The source article's class-specific time choices gave the same stability counts as the 1 s model. A stationary crossing table identified cell 79062 as the smallest positive candidate. It required a gain increase of 16.7% at recurrent eigenvalue 0.99 and 198.4% at 0.9. Dense spectrum checks found earlier oscillatory crossings in two other high-gain paths. Thus the table is not a full stability bound. The models differ in second decay times and are not equal fits to physiology. These are conditional model results, not measured gain or failure distributions in animals. The package preserves code, provenance, every simulation result, checks, figures and repeat instructions. Original source files are referenced by commit and hash.

zebrafish · connectome · neural integrator · neuron gain · stability · intrinsic time · recurrent dynamics