Population-scale atlas of aggregation-propensity perturbation across thirty amyloidogenic disease genes: training data, code, and model checkpoint

Training-reproducibility package for the central statistical result of an unpublished manuscript currently under peer review: a gene-stratified, five-fold cross-validated comparison of AlphaMissense alone against three variants that add structural/aggregation features, across 4,926 missense variants in thirty amyloidogenic disease genes. Contains the exact feature matrix (data/features_ready.csv), the training and cross-validation code (code/reproduce_table4_figure4.ipynb) , and the trained V1model checkpoint (model_v1_honest.pkl), a LightGBM classifier combining AlphaMissense eference performancevalues to check reproduction against are given in the manuscript itself and are not duplicated here.