Weight shuffles gave less mean code change after one input loss in three fly calyx models

We tested whether balanced input to Kenyon cells protects a sparse odor code after one glomerular channel is lost. We linked a verified PN-to-KC edge table from three adult fly calyces to 23 measured receptor channels for 110 odors. FlyWire left and right came from one fly; hemibrain right came from a second fly. Each model selected its strongest 10 percent of cells before and after loss. Observed weights gave mean active-set replacement of 6.74, 6.63 and 6.75 percent. The means of 99 weight shuffles were 6.55, 6.56 and 6.62 percent. Restoring each cell total after a shuffle gave no common protective effect. Observed weights had lower baseline odor-code overlap than shuffle means, but no repeated gain in code preservation. Loss of the worst channel for each odor replaced 44.7, 43.3 and 42.1 percent of active cells on average, despite a fixed active-cell count. These are model stress scores, not measured cell responses or failure probabilities. Different anatomy and receptor animals, missing receptor profiles, count-based weights, a rank threshold and strong selected perturbations limit the result. Code, source hashes, model inputs, all draw scores, paired controls, audit results and figures support reproduction.

Drosophila · mushroom body · Kenyon cells · structure-function · input loss · sparse coding · connectomics · representation reliability · null result