Not the protein but the surface sets where corneal deposits form and stop: model code and digitised source data

Abstract

Model code, digitised source data and figure scripts for a forward model of deposit formation in granular corneal dystrophy type 2 (TGFBI R124H), treated as precipitation of a slightly supersaturated protein in a curved, partly non-renewing tissue rather than as a protein-folding disease. The work has no experimental data of its own. Every clinical value it is tested against is published, and every such value used is included here under data/digitised/ with its source in the file header, so that the reading itself can be checked. The central claim is reproduced by src/_arrest_exponent.py from those two digitised files and nothing else: writing the deposit radius as a ~ t^n and granting the published erosion budget in full, the median admissible exponent across all fifteen pairs of the six published ages is -0.03; one pair admits n > 1/2 and none admits n = 1. No fitted constant, no assumed observation error and no commitment to a growth law enters that comparison. A held-out test is reproduced by src/_mouse_incidence.py against a mouse the model was never fitted to: two constants fixed on the heterozygote incidence curve determine the whole homozygote curve with no free parameter, and it lands within one standard error at all five reported ages. Its expectations were pre-registered before the calculation, in prereg/PREREG_mouse.md. Negative results are kept deliberately (see src/_kappa_lit.py and the notes under prereg/). Three scripts additionally require public single-cell datasets that are too large to deposit; each names the file it needs and where to obtain it. Related patent application: KR 10-2026-0171578, filed 9 September 2026. v1.0.1. Two scripts are new and they carry the deposit's main new result: the ratio of two genotypes' rate constants is not a constant but a decreasing function of the supersaturation itself, r(S) = (rho*S - 1)/(S - 1), so reading that ratio off two published incidence curves bounds the supersaturation. The mutant-concentration ratio rho = 2.04 comes from the source's own protein measurement and is not a free choice. src/_mouse_S0.py performs the measurement; src/_propagate_S0.py carries the bound through the rest of the model. src/_arrest_exponent.py and src/_mouse_incidence.py reproduce the two claims the paper leads with from the digitised files alone. src/figures/figstyle.py gives every figure one palette and type scale. Three corrections. src/_chi_derive.py applied a no-flux boundary condition to the plane source and a zero-concentration condition to the volume source, which cannot be superposed; with them made consistent the local epithelial share falls from 0.950 to 0.619, and the whole-thickness value turns out to be an identity that returns the production split whatever the constants. data/digitised/hong2011_depths.csv repeated the lattice upper-surface standard deviation in the lower-surface column; the correct value is 71.4, which every script and the paper already used. prereg/PREREG_mouse.md carries a second post-hoc note recording that the dosage rule registered before the calculation was the asymptotic case of the correct relation rather than a separate rule.