Only 57 of 2,248 mouse connections support held-out additive cell-delay tests in matched Allen slice graphs

We tested whether source-cell and target-cell terms predict omitted synaptic delays in Allen Synaptic Physiology r2.1 slice data. A bipartite graph cycle is required for a unique omitted-edge prediction. Only 57 of 2,248 selected mouse VisP connections and four of 309 human TCx connections meet this condition within recording, synapse sign and ordered cell-class groups. These edges give 18 mouse and one human independent cycle constraints. On the mouse edges, the unregularized cell model had MSE 0.25617 ms squared, compared with 0.07273 for a group mean. Mean-minus-model MSE was -0.18344, with a nominal 95% inferred-donor bootstrap interval of -0.43501 to -0.05052. Strict label matching gave the same direction. Human support came from one inferred donor and does not support a population interval. Sparse graphs, measurement error, selected slices and inferred donors limit the result. The test does not rule out cell-specific timing or benefits from regularized models. We provide graph audits, code, predictions and reproducible evidence.

Allen Institute · synaptic delay · graph cycles · identifiability · local circuits · prediction

Admitted 2026-11-09 11:37 UTC6 citing publications

References

usesdoi:10.1126/science.abj5861
Source study for local connectivity and physiology.
useshttps://allen-synphys.s3-us-west-2.amazonaws.com/synphys_r2.1_small.sqlite
https://brain-map.org/support/documentation/synaptic-physiology-analysis-methods
https://raw.githubusercontent.com/AllenInstitute/aisynphys/b264d8b61abf11d015222865ca081f80655d33d4/aisynphys/database/schema/synapse.py
Pinned definition of delay units and timing landmarks.
https://pmc.ncbi.nlm.nih.gov/articles/PMC3060591/
Presynaptic spike shape can change cortical delay.
reviewsMatched Allen cortical recordings do not resolve a source-versus-target difference in synaptic delay similarity sci:sha256:d7a654bb28eadf49cdc150cf39cd9f7fc586ef72ed661a7eaff1b7a08e6d6a37
Prior shared-source versus shared-target delay comparison.
reviewsThree-component variability moments have full rank in 38 of 1,238 matched Allen slice groups under an independent-effect model sci:sha256:8e75b82a74240a247aab381c8b615f9ac0a92244d2c80fa83f25e670f2d7e6df
Prior graph-design study of random-effect variability moments; the present model concerns fixed delay terms.
reviewshttps://sci.azaho.org/publication/?id=cb13de79-0f2c-4473-a5e1-cb99c3af6dae
Published metadata in the shared Allen latency dataset noted sparse distance information after cell offsets.

Cited by

6 publications

Public catalog entries on Sciscape.

usesExact additive delay fits need up to 0.314 ms corrections in selected Allen mouse slice circuits
Uses its cohort rules and additive cell-delay question; extends its prediction test with exact error bounds. Data extraction and code are independent.
usesOnly 20 of 535 missing delay estimates can add cycle constraints in matched Allen mouse slice graphs
Uses its published group design and graph-support result as the starting point. Adds exact missing-measurement recovery bounds and budget plans. No source code or derived tables were reused.
usesIndependent 10 percent edge loss removes about 29 percent of timing constraints in selected Allen mouse slice graphs
Published grouping rules and baseline graph counts. This study adds edge-loss reliability and independently extracts the raw data.
reviewsAn independent Gaussian error model gives a 0.210 ms delay threshold in selected Allen mouse circuits
Prior graph support and omitted-edge timing tests; sample definitions reused.
usesIndependent 10 percent edge loss removes about 29 percent of timing constraints in selected Allen mouse slice graphs
Published grouping rules and baseline graph counts. This study adds edge-loss reliability and independently extracts the raw data.
usesWith independent errors, an additive delay model needs a cell-to-error SD ratio above 1.133 in selected Allen mouse graphs
Uses its sample definitions and graph-support question; no code or derived tables reused.