Methods note: synapse density around plaques and amyloid-laden somata in the public Han et al. 3xTg volume

Written October 21, 2026 for an outside reader. Every number below is in results.json, produced by readout.py on the same day; section letters refer to results.md. Where a statement is an assumption rather than a measurement it is labeled as one.

Dataset. Han et al., bioRxiv 2023.10.24.563674 v4 (February 2025), all rights reserved, no license on the bucket. One female one-year-old 3xTg mouse, hippocampal CA1, imaged by multibeam serial-section electron microscopy with nanobody fluorescence for amyloid-beta, phospho-tau, CD11b and Hoechst registered into the EM. The preprint states the imaged volume as 245 x 238 x 23 µm. The synapse annotation layer covers a smaller box, 188.6 x 167.7 x 15.2 µm (0.00048 mm³); all densities here are computed inside that box. Public location: precomputed://gs://vclem-xh/alzheimers/.

Synapse detectors. Two automated layers from the preprint, both used as published. The transfer-learning layer (synapses/20241024) holds 732,791 line annotations with pre- and postsynaptic segment IDs; the preprint states precision 46 percent and recall 52 percent on its evaluation volume (figures taken from our reading of the preprint, not re-measured here). The supervised layer (wei_synapses/syn_20241121) is read as pre/post segment pairs at 64 nm and counted once per pair, 576,643 inside the box; stated precision 86 percent, recall 99 percent. A synapse is placed at its presynaptic point (transfer layer) or its segment centroid (supervised layer). Far-field densities are 1.607 and 1.266 per µm³ respectively; the two layers disagree on absolute counts and agree on every relative value in section A.

Plaque mask. Amyloid-beta channel at 256 x 256 x 240 nm, intensity at least 60 of 255, connected components of at least 20 µm³, closed with two binary-closing iterations. Of 43 components, the two whose centroids lie within 8 µm of the preprint's own plaque coordinates are treated as extracellular plaques (plaque 1, 226 µm³, 2.2 µm from the listed point; plaque 2, 582 µm³, 1.7 µm). The other 41 are treated as intraneuronal amyloid, called somata below; that classification is an assumption inherited from the preprint's description, not a check of each object against the EM. Threshold sensitivity (A3): at 40, 60 and 90 the 0 to 5 µm relative density is 0.36, 0.30 and 0.25.

Distances, denominator, shells, reference. Distance is the Euclidean distance transform from the plaque edge with anisotropic voxel spacing, not from the plaque center. The denominator is tissue volume inside the annotated box and outside plaque cores. It is tissue, not neuropil: somata, nuclei and vessels are not excluded, so the density is synapses per µm³ of everything present. Shells are 0 to 5, 5 to 10, 10 to 15, 15 to 20, 20 to 30, 30 to 50, 50 to 100 and over 100 µm. The far-field reference is tissue at least 50 µm from every plaque. Poisson errors describe counting only. Block-bootstrap intervals resample the box's 3 x 3 grid of xy blocks with replacement, 500 times, and recompute relative density against the resampled far field; with nine blocks and one plaque these intervals are crude, and the two weightings used in the file (A2 combines per-shell relatives with fixed tissue weights; section D recomputes the pooled ratio inside each resample) give 0 to 10 µm intervals of [0.34, 0.43] and [0.21, 0.43] for the same point estimate.

Laminar control. Plaque 1 sits in a cell-body layer; density differs between CA1 laminae on its own, so a gradient away from the plaque can partly be a gradient out of the lamina. The control places virtual centers at the plaque's own y and z, shifted along x (the direction the lamina runs) by -80, -60, -40 and +40 µm where the center stays 25 µm inside the box, over tissue at least 25 µm from every real plaque, and computes the same shells. The lamina-corrected value is the real pooled shell divided by the pooled virtual shell.

Per-plaque results (D, E). Plaque 1 is wholly inside the box. Its 0 to 5 µm shell holds 1,598 transfer synapses in 3,282 µm³, relative density 0.30 [0.27, 0.34]; supervised 0.29 [0.27, 0.31]. Pooled 0 to 10 µm: 0.38 transfer, 0.35 supervised. The 15 µm slab truncates its neighborhood: 61 percent of the geometric 0 to 10 µm shell and 22 percent of the 0 to 30 µm shell lie inside the box. Its virtual lamina runs at 0.67 of the far field on the transfer layer, giving a corrected 0 to 10 µm value of 0.57 [0.31, 0.64] and 0.72 at 10 to 20 µm; on the supervised layer the lamina runs at 0.90, giving 0.38 [0.26, 0.41] and 0.58. The two detectors therefore agree on the raw gradient and disagree on how much of it the lamina explains. Plaque 2 lies entirely outside the annotated box: its centroid is 12.9 µm beyond the lower y face and 5.2 µm beyond the upper z face, none of its mask and 0 percent of its 0 to 10 µm shell fall inside, and only 6 percent of its 0 to 30 µm shell does. Its observable shells, 10 to 20 µm with 3,502 µm³ of tissue, are at 0.87 [0.78, 1.01] transfer and 0.91 [0.84, 0.98] supervised, in a lamina the control puts at 0.71 and 0.93 on 653 µm³ of virtual tissue. Nothing can be said about plaque 2 within 10 µm. The pooled section A profile is therefore plaque 1's profile with a small contribution beyond 10 µm from plaque 2.

Per-soma results (F). Each of the 41 objects is profiled on its own, distance from its edge, with the interiors of all somata excluded and tissue within 20 µm of a plaque removed, relative to the section B reference of the same detector (1.568 transfer, 1.285 supervised). 28 objects have some volume inside the box and 31 have at least 100 µm³ of tissue in their 0 to 5 µm shell; the rest sit in the 60 µm margin outside the box and are listed but not summarized. Over those 31, the transfer layer gives a median 0 to 5 µm relative density of 0.91, interquartile range 0.71 to 1.05, range 0.38 to 1.57, with 19 below 1; the supervised layer gives 0.83, 0.64 to 0.98, range 0.35 to 1.31, with 24 below 1. At 5 to 10 µm both medians are 0.96 to 0.97. The pooled section B value is 0.87 (transfer) and 0.79 (supervised) in the 0 to 5 µm shell. Poisson intervals exclude 1 for nearly every soma below 1, which reflects large counts, not independent evidence; the spread across somata is the honest measure.

Bulk and ratio figures. The bulk deficit is the overall box density, 1.506, against the far field, 1.607: 6.2 percent. The 5.5 percent is the fraction of box tissue within 20 µm of a plaque edge. The sevenfold figure is the lamina-corrected 0 to 10 µm deficit on the transfer layer, 43 percent (1 minus 0.57), divided by 6.2. On the supervised layer the same ratio would be 62 divided by 6.2, about ten. Section G gives the deficit a field of view centered on plaque 1 would report: 52 percent within 20 µm, 20 percent within 100 µm (transfer).

Comparison with Han et al. The preprint reports 0.181, 0.653 and 1.216 synapses per µm³ in nine hand-placed 6 µm boxes at 10, 20 and 30 µm from the plaque center, an 85 percent drop from 30 to 10 µm. The same bands around plaque 1's centroid here give 0.629, 0.857 and 1.163. The analyses differ in that the preprint measures from the center and this one from the edge, and the preprint samples boxes along chosen directions while whole shells average over every direction including the lamina, where the deficit is smaller (section A1b: 0.27 at 0 to 10 µm along +x, 0.37 along -x).

What this does not establish. One animal, one region, no non-transgenic control, no treatment, no between-animal variance. The synapse counts are automated predictions at the stated precision and recall; nothing was checked against the EM by hand. The cohort table in results.md is arithmetic under assumed effect and variance, not a power analysis. The lamina correction depends on which detector is believed.

Reproduce. About 1 GB of public data plus the supervised layer (about 15 minutes to extract). With uv installed, from experiments/synaptic-readout: set READOUT_DATA, run the fetch scripts in the README, then uv run --python 3.12 --with numpy --with scipy --with pandas python readout.py. Runtime about three minutes. config.json holds every threshold and the plaque coordinates.