A 128 ms window shift changed a depth model's error decrease from -6.4% to 20.1% in 178 cells from a second mouse retina

Dendrite depth gives a way to predict light-on and light-off responses from retinal anatomy. We tested transfer between two mouse retina datasets. We used 315 e2198 ganglion cells and 178 Eyewire II ganglion cells. The target was a local calcium edge index from moving-bar traces. Models used ON-side profile mass or ten depth bins. Each test retina was excluded from training and penalty selection. In the primary test, the depth model trained on e2198 increased error by 6.4% relative to the training mean in Eyewire II. Moving all local baseline and response samples 128 ms later gave a 20.1% error decrease. Conditional type ranges were -42.1% to 27.7% and -14.7% to 48.8%, respectively. Reverse transfer error decreases were 14.8% and 34.7%. An exploratory check kept the original model fixed. Moving only the test windows later changed its error decrease to 24.4%. The primary index had repeat-group correlations of 0.928 and 0.902, but errors differed between recording fields. Thus, repeat agreement and aligned anatomy did not make this structure-function score stable. These results describe a declared calcium index in two tissues. They do not select the correct biological window or identify the cause of the timing effect. The package gives code, source hashes, aggregate results, figures, and a complete repeat procedure.

retina · structure-function · external validation · calcium imaging · dendrite depth · response timing · reproducibility