Luoshu Pocket Geometry Predicts Protein Druggability: Pocket Shape and Depth as Orthogonal Geometric Signals (v3, n=85)

v3 (ablation-validated): Two orthogonal geometric signals from the Luoshu framework predict protein druggability with AUC=0.851 on 85 DUD-E proteins (perm-p<0.001), exceeding SiteMap (~0.74) by +0.111. Signal 1: pocket sphericity (AUC=0.814) -- druggable pockets are sphere-like; undruggable PPI surfaces are flat. Signal 2: pocket depth chc_max (AUC=0.781) -- deep pockets are druggable. Optimal composite: 0.7*sphericity + 0.3*chc_max (AUC=0.851). Ablation confirms sphericity dominates (removal: AUC 0.851->0.781) while chc_max adds complementary information (removal: 0.851->0.814). Head-to-head vs pseudo-fpocket (same 85 proteins): +0.103. Method: C-alpha + backbone dihedrals only, <1 second, no ML/MD/ligand data. Transparency: v1 (n=24, AUC=0.957 proof-of-concept), v2 (n=85, AUC=0.846 phi-weighted), v3 (n=85, AUC=0.851 ablation-validated). Includes complete data table (druggability_v4_table.csv, n=85). Extends: doi:10.5281/zenodo.20073888 (v2), doi:10.5281/zenodo.20073303