Common partner counts gave no clear connection prediction gain in selected Allen mouse and human slices

We tested whether common input and output partners improve prediction of a detected chemical connection after other measured wiring features are known. We used Allen Synaptic Physiology r2.1 and excluded whole inferred donors from fitting. The samples had 18,906 mouse VisP directions from 732 inferred donors and 1,456 human TCx directions from 42 inferred donors. Base models used cell labels, distance, probe counts, the reverse call, local marginal connection counts, and recording density. Adding common partner counts gave log-loss gains of -0.000100 and -0.001553 nats per direction. Conditional 97.5% donor intervals were [-0.000371, 0.000161] and [-0.003547, 0.000404]. Neither primary test gave a clear gain. With local marginal counts fixed, either overlap count could vary in only 8.65% of mouse directions and 12.64% of human directions. This is a conditional completion benchmark, not a test of unmeasured whole circuits or neural activity. Sparse overlap, slice detection, inferred donor labels, model choice, and intervals that omit refit uncertainty limit inference. We supply selected calls, predictions, overlap bounds, source hashes, code, figures, and an independent audit.

Allen Institute · common neighbor rule · connection prediction · donor validation · cortical slices · null result · overlap bounds