HEDGEHOG: hierarchical evaluation of drug generators through rigorous filtration.

HEDGEHOG applied to a KRAS G12D de novo drug-generation case study found final survival across 22 generative models ranged from zero to 127.0 ± 18.1 molecules per run.

• Why it matters: Standard generator metrics weakly indicate whether candidates are medicinally plausible and computationally suitable, risking misleading performance views and wasted downstream effort.
• What they did: The authors built HEDGEHOG, a configurable six-stage hit-identification workflow (descriptor screening, structural alerts/graph checks, synthesis feasibility, docking, and 3D pose/interaction checks) and ran three sets of 1000 generation attempts per model.
• The result: After multiple-testing correction, standard metrics showed no significant association with final survival, while HEDGEHOG enabled reproducible, stage-wise failure-mode identification to guide generator development for early discovery.

AI summary of the abstract · may contain errors · not medical advice · read the paper.