HEDGEHOG Benchmark Filters Drug Candidates for Plausibility

Generative molecular models are increasingly used in early drug discovery to propose new candidate compounds. However, current evaluation metrics for these models often fail to reflect whether the generated molecules are truly medicinally plausible, synthetically accessible, or suitable for downstream computational analysis. This can lead to false positives and inefficient resource use. To address this, researchers developed HEDGEHOG, a comprehensive six-stage filtration benchmark. Inspired by industrial hit identification workflows, HEDGEHOG includes rigorous checks for physicochemical properties, structural alerts, synthesis feasibility, docking and binding affinity, and 3D pose interactions. Evaluating 23 molecular generators across three model classes and 230,000 generated molecules, HEDGEHOG revealed a stark reality: only 0.65% of the initial molecules survived all filtration stages. This finding underscores a critical limitation of current generative models, which struggle to simultaneously satisfy the multi-parameter design constraints required for viable drug candidates.