Cheminformatics

HEDGEHOG: Hierarchical Evaluation of Drug Generators Through Rigorous Filtration

Daria A. Ryabchenko, Pavel Gurevich, Shamil Kadyrov, Daria Frolova, Kseniia Fedisheva, Sergei A. Nikolenko, Alexander Shapeev, Marina A. Pak

Featured July 18, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Evaluating new drug-making AI models with a tough, step-by-step filter, like a real drug lab, shows that most generated molecules fail basic checks, revealing a big gap between what AI makes and what real drugs need.

In depth

The paper introduces HEDGEHOG, a novel six-stage hierarchical filtration benchmark designed to rigorously evaluate generative molecular models. Unlike traditional metrics that often overestimate utility, HEDGEHOG simulates a realistic drug discovery workflow, applying sequential filters for preprocessing, physicochemical properties, structural alerts, synthesis feasibility, docking, and 3D pose checks. This approach reveals that only a tiny fraction of generated molecules (0.65%) simultaneously satisfy all medicinal chemistry and target-aware constraints, exposing a critical limitation of current generators.

Key Takeaways

1
Current molecular generator evaluation metrics often lead to false positives, as they do not adequately reflect the multi-parameter constraints of practical drug discovery workflows.
2
The HEDGEHOG benchmark introduces a six-stage, coarse-to-fine filtration cascade that mimics industrial hit identification, including preprocessing, physicochemical screening, structural alerts, synthesis feasibility, docking, and 3D pose checks.
3
The study demonstrates that only a very small percentage (0.65%) of generated molecules survive all stages, highlighting that models struggle to produce compounds that simultaneously satisfy chemical plausibility, synthetic tractability, and target compatibility.

Conceptual Flow

HIGH LEVEL

Methodology: Multi-Stage Drug Candidate Filtering

The new method checks potential drug molecules through six strict steps, just like real scientists would, to find only the very best ones.

New Molecules

Pass Through Filters

Cleaned Molecules

Good Properties

Safe Structures

Easy to Make

Bind Well

Good Shape

Results: High Attrition, Low Survival

Most new molecules fail early checks, and only a tiny fraction make it through all the steps, showing that making good drugs is very hard.

Many New Molecules

Fail Most Checks

Few Good Molecules