Comparator-Dependent Druggability and Safety of Spiro-Terpenoid Scaffolds: A Proteome-Scale Pocket-Transfer Screen

Abstract

Proteome-scale screening is attractive for polypharmacology and safety assessment but is often dominated by computationally expensive docking and opaque target-prediction models. We developed a ligand-conditioned pocket-transfer workflow that combines public chemical structures, experimentally observed template ligands, and precomputed human pocket similarities. A new 60-compound dataset comprised 23 spiro-terpenoids and 37 structured controls. Morgan-fingerprint similarity, APoc pocket similarity, one bounded propagation step, 41 ADMET-AI endpoints, and a four-axis weight-robust analysis generated 1,215,300 compound–protein scores across 20,255 human proteins. Among 36 benchmark drugs with evaluable targets, the median first-target rank was 10.5; macro recall at ranks 100 and 500 was 0.148 and 0.217, or 30.0- and 8.80-fold random enrichment. NR3C2, the known eplerenone target, was unreachable and ranked 20,256, exposing a decisive coverage failure. Spiro-terpenoids showed no universal target-opportunity or composite-safety advantage over nonspiro terpenoids, whereas predicted liabilities were lower than for planar-drug controls. Khusimone combined Fsp3 0.786 with predicted hERG and DILI probabilities of 0.114 and 0.099; spirojatamol combined Fsp3 0.867 with low predicted DILI (0.028) but higher hERG probability (0.288), illustrating compound-specific trade-offs. HSD17B1, ESR2, and PXR hypotheses define a minimal orthogonal testing panel. The CPU-light, fully reproducible scores prioritize experiments rather than estimate binding affinity or clinical safety.