HEDGEHOG: Hierarchical Evaluation of Drug Generators Through Rigorous Filtration

arXiv:2607.13155v2 Announce Type: replace Abstract: Generative molecular models can support early drug discovery by proposing new candidate compounds de novo. In practice, useful candidates must balance target-relevant activity, physicochemical properties, and other multiparameter design constraints. However, standard metrics commonly used to evaluate molecular generators only weakly reflect whether the generated compounds are medicinally plausible and suitable for downstream computation. This can produce an incomplete view of model performance and inefficient use of computational resources. We introduce HEDGEHOG, a unified six-stage filtration benchmark that is constructed as a hit identification workflow: (i) preprocessing; (ii) physicochemical descriptor screening; (iii) structural alerts and graph-sanity checks; (iv) synthesis feasibility; (v) docking; and (vi) three-dimensional pose and interaction checks. We evaluated 22 generative models in a KRAS G12D case study, using three runs of 1,000 requested generation attempts per model. The models showed different patterns of attrition, and final survival ranged from 0 to 127 molecules per run. None of the standard metrics showed a significant association with final survival after correction for multiple testing. HEDGEHOG provides a reproducible benchmark for evaluating molecular generative models by measuring molecule survival through chemical filters. The framework identifies stage-wise failure modes across generator classes and provides a practical basis for developing molecular generators better aligned with early drug discovery.

Authors:Daria A. Ryabchenko (1,2), Pavel Gurevich (1,2), Shamil Kadyrov (1), Daria Frolova (1,2), Kseniia Fedisheva (1), Sergei A. Nikolenko (1), Alexander Shapeev (1,2), Marina A. Pak (1) ((1) Ligand Pro, Moscow, Russia, (2) Skolkovo Institute of Science and Technology, Artificial Intelligence Center, Moscow, Russia)

Abstract:Generative molecular models can support early drug discovery by proposing new candidate compounds de novo. In practice, useful candidates must balance target-relevant activity, physicochemical properties, and other multiparameter design constraints. However, standard metrics commonly used to evaluate molecular generators only weakly reflect whether the generated compounds are medicinally plausible and suitable for downstream computation. This can produce an incomplete view of model performance and inefficient use of computational resources. We introduce HEDGEHOG, a unified six-stage filtration benchmark that is constructed as a hit identification workflow: (i) preprocessing; (ii) physicochemical descriptor screening; (iii) structural alerts and graph-sanity checks; (iv) synthesis feasibility; (v) docking; and (vi) three-dimensional pose and interaction checks. We evaluated 22 generative models in a KRAS G12D case study, using three runs of 1,000 requested generation attempts per model. The models showed different patterns of attrition, and final survival ranged from 0 to 127 molecules per run. None of the standard metrics showed a significant association with final survival after correction for multiple testing. HEDGEHOG provides a reproducible benchmark for evaluating molecular generative models by measuring molecule survival through chemical filters. The framework identifies stage-wise failure modes across generator classes and provides a practical basis for developing molecular generators better aligned with early drug discovery.

Machine Learning (cs.LG); Software Engineering (cs.SE)

arXiv:2607.13155 [cs.LG]

(or arXiv:2607.13155v2 [cs.LG] for this version)

From: Daria A. Ryabchenko [view email] [v1] Tue, 14 Jul 2026 18:05:28 UTC (2,428 KB) [v2] Tue, 29 Sep 2026 09:29:55 UTC (3,291 KB)

aiqbench analysis

HEDGEHOG: Hierarchical Evaluation of Drug Generators Through Rigorous Filtration

Related coverage