With donor groups kept separate, soma distance gave 3–12% lower delay test error after cell label controls in three cortical samples

We tested whether soma distance improves synaptic delay prediction after cell label controls. We used 814 mouse and 130 human connections from Allen Synaptic Physiology r2.1, plus 954 human connections from a separate fixed release. Nested ridge tests kept complete donor groups separate and gave each group equal weight. In the first fixed split, adding linear distance reduced mean squared test error by 11.5%, 9.3%, and 2.95%. The 95% donor resample ranges were −2.4% to 20.6%, 0.3% to 18.5%, and −0.31% to 6.31%. These ranges hold predictions fixed. With distance at most 300 micrometres, gains were 5.4%, 4.8%, and 2.7%. A few longer Allen links supplied about 40% of the full-sample error reduction. All three class-adjusted distance slopes were positive, but combined model root mean squared errors remained 0.57–0.62 milliseconds. A wider penalty grid gave full-sample gains of 11.5%, 10.3%, and 2.0%. Geometry added timing information but left much individual error under these models. Slice selection, incomplete labels, inferred Allen donors, and incomplete training uncertainty limit inference. The package gives source hashes, code, cohorts, predictions, losses, plans, figures, and repeat instructions.

synaptic delay · structure and function · soma distance · cortical circuits · donor validation · paired recordings